ErrOPD: When Does On-Policy Distillation Transfer Teacher Errors?

arXiv:2605.08737 · cs.LG, cs.CL · Submitted 2026-05-09 · Read on arXiv

cs.LG, cs.CL

Submitted: 2026-05-09

Updated: 2026-10-12

Comments: Project page: https://lixin.ai/ErrOPD

Project page: https://lixin.ai/ListOPD

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Terminology

• Distillation — 2× in this paper · explained in 2 episodes
• On-Policy Distillation — 2× in this paper · explained in 3 episodes

Sources

• Think Inside the JSON: Reinforcement Strategy for Strict LLM Schema Adherence
• MS MARCO: A Human Generated MAchine Reading COmprehension Dataset
• Guiding LLMs The Right Way: Fast, Non-Invasive Constrained Generation
• Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities
• Overview of the TREC 2020 deep learning track
• Decoupling Task-Solving and Output Formatting in LLM Generation
• Revisiting On-Policy Distillation: Empirical Failure Modes and Simple Fixes
• JSONSchemaBench: A Rigorous Benchmark of Structured Outputs for Language Models
• Distilling the Knowledge in a Neural Network
• Bridging Language and Items for Retrieval and Recommendation: Benchmarking LLMs as Semantic Encoders
• Stable On-Policy Distillation through Adaptive Target Reformulation
• RLPO: Residual Listwise Preference Optimization for Long-Context Review Ranking
• Entropy-Aware On-Policy Distillation of Language Models
• Coverage Improvement and Fast Convergence of On-policy Preference Learning
• DistiLLM: Towards Streamlined Distillation for Large Language Models
• DistiLLM-2: A Contrastive Approach Boosts the Distillation of LLMs
• Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe
• Zero-Shot Listwise Document Reranking with a Large Language Model
• RankZephyr: Effective and Robust Zero-Shot Listwise Reranking is a Breeze!
• The Hidden Cost of Structured Generation in LLMs: Draft-Conditioned Constrained Decoding

Related papers

• Polynomial-Augmented Neural Networks (PANNs) with Weak Orthogonality Constraints for Enhanced Function and PDE Approximation
• AIRL-S: Unifying Reinforcement Learning and Search-Based Test-Time Scaling via Adversarial Inverse Reinforcement Learning
• Transformers as Bayesian In-Context Experimenters: Smoothness-Adaptive Efficient ATE Estimation
• Convergence issues in Relational Concept Analysis based on AOC-posets
• Beliefs Beyond Posteriors: The Predictive Cost of ELBO Binding
• Understanding Diffusion Models via Ratio-Based Function Approximation with SignReLU Networks